[extropy-chat] 70 teraflops

[Eugen Leitl]

On Fri, Dec 03, 2004 at 12:59:28PM +0000, Dirk Bruere wrote:

> Depends.
> In the past reducing scale also meant increasing clock speed.

Clock speed is not system performance. Right now there are some severe speed
bumps associated with structure shrink (leak currents) and heat dissipation.
Successor technologies (carbon nanotube transistors) promise to solve that,
but are currently lab unicates. It's not obvious you can grow useful arrays
out of gas phase, and you sure can't put them in place individually. Self
assembly doesn't seem to help here.

> Now it seems that processors are going parallel (something long overdue 

They're not. The only COTS parallelism is at the CPU level (multicore, which
is idiotic, since memory-starved) and individual system level (lousy latency,
lousy bandwidth with current signalling mesh -- some improvements with
Infiniband and 10 GBit direct memory access NICs, which give you ~us userland
latency). 

None of this is relevant for typical nonscientific code. MPI is great, but
nobody is using it outside of scientific supercomputing. No good parallel
debuggers, either. And of course no parallel programmers, no code base, no
nothing. Does it suck? And how.

If you want to build affordable parallel systems with good performance, you
have to put multiple cores with embedded memory and crossbars as well as
signalling interconnect on die. You can't get conventional OS'ses into such tight
memory spaces (though some very small practical systems exist both in
industry and academia), and you sure can't build vanilla open source stuff on
such highly parallel systems with tiny cores. All off-die memory access will be via
signalling mesh, and hence expensive.

If you could buy them, that is. Right now resources are only available during
periods of positive cash flow, and are immediately dumped into new foundries,
which need to be run at full load producing cash chips. Radical R&D is not
available in the current climate. Things will only improve if you can scale
down the costs of building and running the fabs.

> IMO) and clock speed is not going to increase as rapidly as in the past.
> Hence Moore's Law will be very closely linked to computing power.

You're projecting our desires again. The reality is different, both the past,
the present, and the near future.

What will happen after that is not yet obvious. 

-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07078, 11.61144            http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
http://moleculardevices.org         http://nanomachines.net
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20041203/83dd8b94/attachment.bin>